Towards Learning Causal Representations of Technical Word Embeddings for Smart Troubleshooting

Abstract

This work explores how the causality inference paradigm may be applied to troubleshoot the root causes of failures through language processing and Deep Learning. To do so, the causality hierarchy has been taken for reference: associative, interventional, and retrospective levels of causality have thus been researched within textual data in the form of a failure analysis ontology and a set of written records on Return On Experience. A novel approach to extracting linguistic knowledge has been devised through the joint embedding of two contextualized Bag-Of-Words models, which defines both a probabilistic framework and a distributed representation of the underlying causal semantics. This method has been applied to the maintenance of rolling stock bogies, and the results indicate that the inference of causality has been partially attained with the currently available technical documentation (consensus over 70%). However, there is still some disagreement between root causes and problems that leads to confusion and uncertainty. In consequence, the proposed approach may be used as a strategy to detect lexical imprecision, make writing recommendations in the form of standard reporting guidelines, and ultimately help produce clearer diagnosis materials to increase the safety of the railway service.

Evaluating the Performance of ChatGPT in the Automation of Maintenance Recommendations for Prognostics and Health Management

“…Pau, Tarquini, Iannitelli, and Allegorico (2021) (Pau, Tarquini, Iannitelli, & Allegorico, 2021) utilized NLP tech-niques for consistent troubleshooting insights in an M&D center, while Peshave et al (Peshave et al, 2022) evaluated approaches for vectorization of short-text case titles. Trilla, Mijatovic and Vilasis-Cardona (2022) (Trilla, Mijatovic, & Vilasis-Cardona, 2022) used TLP for troubleshooting in PHM and developed a failure ontology and a data-driven quality strategy. Pires, Leitao, Moreira and Ahmad (2023) (Pires, Leitão, Moreira, & Ahmad, 2023) compared different recommendation systems, including their own discrete event simulation model, and demonstrated improved user ratings over state-of-the-art recommendation systems.…”